fSGLD ICML-2026 Reproduction (Kw0nhxeTeI)
Collection
Reproduction of Flatness-Aware SGLD (arXiv 2510.02174, ICML 2026): theory + CIFAR-100N experiments. • 2 items • Updated
args dict | num_classes int64 100 100 | best_test_acc float64 10.1 73.7 | final_test_acc float64 10.1 72.4 | avg_last10_test_acc float64 8.94 71.9 | s_per_epoch float64 29.1 784 | total_time_s float64 81 9.75k | history listlengths 2 150 | hessian dict |
|---|---|---|---|---|---|---|---|---|
{
"dataset": "cifar100N",
"backbone": "resnet34",
"optimizer": "fsgld",
"epochs": 2,
"milestones": [
50,
100
],
"batch_size": 128,
"lr": 0.1,
"sigma": 0.01,
"beta_coupling": true,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 8,
"data_root": "./data",
"tag": "_canary",
"repo... | 100 | 10.07 | 10.07 | 8.935 | 40.502335 | 81.005643 | [
{
"epoch": 0,
"train_loss": 4.38526275273784,
"train_acc": 4.78,
"test_loss": 4.059505396251437,
"test_acc": 7.8,
"time": 40.75534510612488
},
{
"epoch": 1,
"train_loss": 4.021495728846401,
"train_acc": 9.222,
"test_loss": 3.857208662395236,
"test_acc": 10.07,
"ti... | {
"lambda_max": 207.40167236328125,
"hessian_trace": 1598.1996276855468,
"trace_se": 139.96819598742047,
"n_params": 21328292,
"n_batches": 5,
"power_iters": 25,
"hutch_probes": 30
} |
{
"dataset": "cifar100N",
"backbone": "resnet34",
"optimizer": "fsgld",
"epochs": 150,
"milestones": [
50,
100
],
"batch_size": 128,
"lr": 0.1,
"sigma": 0.003,
"beta_coupling": true,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 8,
"data_root": "./data",
"tag": "_coupled_s003"... | 100 | 53.72 | 53.21 | 53.255 | 40.310397 | 6,048.092113 | [
{
"epoch": 0,
"train_loss": 4.075463124858144,
"train_acc": 9.912,
"test_loss": 3.6160612709914584,
"test_acc": 14.83,
"time": 40.8138473033905
},
{
"epoch": 1,
"train_loss": 3.633580847469437,
"train_acc": 16.792,
"test_loss": 3.2635345941857445,
"test_acc": 20.53,
... | {
"lambda_max": 134.36517333984375,
"hessian_trace": 1960.287373860677,
"trace_se": 48.88110694405228,
"n_params": 21328292,
"n_batches": 5,
"power_iters": 25,
"hutch_probes": 30
} |
{
"dataset": "cifar100N",
"backbone": "resnet34",
"optimizer": "fsgld",
"epochs": 150,
"milestones": [
50,
100
],
"batch_size": 128,
"lr": 0.1,
"sigma": 0.01,
"beta_coupling": true,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 8,
"data_root": "./data",
"tag": "_coupled",
"r... | 100 | 59.88 | 57.36 | 57.848 | 40.280189 | 6,041.953324 | [
{
"epoch": 0,
"train_loss": 4.329296830365115,
"train_acc": 5.714,
"test_loss": 4.042809625215169,
"test_acc": 8.04,
"time": 40.773048877716064
},
{
"epoch": 1,
"train_loss": 4.002977585853518,
"train_acc": 9.716,
"test_loss": 3.81444269795961,
"test_acc": 10.64,
... | {
"lambda_max": 269.2916564941406,
"hessian_trace": 8908.325341796875,
"trace_se": 202.89016716000285,
"n_params": 21328292,
"n_batches": 5,
"power_iters": 25,
"hutch_probes": 30
} |
{
"dataset": "cifar100N",
"backbone": "resnet34",
"optimizer": "fsgld",
"epochs": 150,
"milestones": [
50,
100
],
"batch_size": 128,
"lr": 0.1,
"sigma": 0.01,
"beta_coupling": false,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 8,
"data_root": "./data",
"tag": "_fixed",
"re... | 100 | 59.21 | 57.33 | 57.337 | 40.298473 | 6,045.639057 | [
{
"epoch": 0,
"train_loss": 4.3293989838846505,
"train_acc": 5.726,
"test_loss": 4.0689939607547805,
"test_acc": 7.17,
"time": 40.80978751182556
},
{
"epoch": 1,
"train_loss": 4.015034605475033,
"train_acc": 9.558,
"test_loss": 3.827103207382975,
"test_acc": 10.92,
... | {
"lambda_max": 294.328125,
"hessian_trace": 8895.93046875,
"trace_se": 278.11761724543646,
"n_params": 21328292,
"n_batches": 5,
"power_iters": 25,
"hutch_probes": 30
} |
{
"dataset": "cifar100N",
"backbone": "resnet34",
"optimizer": "sam",
"epochs": 150,
"milestones": [
50,
100
],
"batch_size": 128,
"lr": 0.1,
"sigma": 0.01,
"beta_coupling": false,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 8,
"data_root": "./data",
"tag": "",
"repo": "tx... | 100 | 62.39 | 56.16 | 56.328 | 65.01844 | 9,754.348295 | [
{
"epoch": 0,
"train_loss": 4.407560836933458,
"train_acc": 5.174,
"test_loss": 4.041115863413751,
"test_acc": 7.18,
"time": 65.53149509429932
},
{
"epoch": 1,
"train_loss": 4.020354398985958,
"train_acc": 9.528,
"test_loss": 3.7766267860992047,
"test_acc": 11.29,
... | {
"lambda_max": 15.207082748413086,
"hessian_trace": 515.7875946044921,
"trace_se": 11.07386574968274,
"n_params": 21328292,
"n_batches": 5,
"power_iters": 25,
"hutch_probes": 30
} |
{
"dataset": "cifar100N",
"backbone": "resnet34",
"optimizer": "sgd",
"epochs": 150,
"milestones": [
50,
100
],
"batch_size": 128,
"lr": 0.1,
"sigma": 0.01,
"beta_coupling": false,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 8,
"data_root": "./data",
"tag": "",
"repo": "tx... | 100 | 60.92 | 54.36 | 54.31 | 29.148646 | 4,373.127709 | [
{
"epoch": 0,
"train_loss": 4.348543735416344,
"train_acc": 4.828,
"test_loss": 4.001657537267178,
"test_acc": 8.03,
"time": 29.62450933456421
},
{
"epoch": 1,
"train_loss": 3.9408378765711087,
"train_acc": 9.452,
"test_loss": 3.715222111231164,
"test_acc": 12.14,
... | {
"lambda_max": 53.519874572753906,
"hessian_trace": 684.7972930908203,
"trace_se": 29.171989493270637,
"n_params": 21328292,
"n_batches": 5,
"power_iters": 25,
"hutch_probes": 30
} |
{
"dataset": "cifar100N",
"backbone": "resnet34",
"optimizer": "sgld",
"epochs": 150,
"milestones": [
50,
100
],
"batch_size": 128,
"lr": 0.1,
"sigma": 0.01,
"beta_coupling": false,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 8,
"data_root": "./data",
"tag": "",
"repo": "t... | 100 | 62.07 | 55.15 | 55.279 | 37.918362 | 5,689.862642 | [
{
"epoch": 0,
"train_loss": 4.281171106621432,
"train_acc": 6.622,
"test_loss": 3.9353790071946158,
"test_acc": 9.5,
"time": 38.47290825843811
},
{
"epoch": 1,
"train_loss": 3.8711220157115966,
"train_acc": 11.866,
"test_loss": 3.6183183676079858,
"test_acc": 13.75,
... | {
"lambda_max": 24.933250427246094,
"hessian_trace": 483.9316884358724,
"trace_se": 12.251427814172203,
"n_params": 21328292,
"n_batches": 5,
"power_iters": 25,
"hutch_probes": 30
} |
{
"dataset": "cifar100N",
"backbone": "vit_b_16",
"optimizer": "fsgld",
"epochs": 10,
"milestones": [
7
],
"batch_size": 128,
"lr": 0.01,
"sigma": 0.005,
"beta_coupling": true,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 12,
"data_root": "./data",
"tag": "_coupled",
"repo": "t... | 100 | 56.99 | 56.99 | 43.301 | 456.946303 | 4,572.57352 | [
{
"epoch": 0,
"train_loss": 4.501770987839955,
"train_acc": 4.058,
"test_loss": 4.301164295099959,
"test_acc": 9.31,
"time": 458.52096915245056
},
{
"epoch": 1,
"train_loss": 4.110906125334522,
"train_acc": 16.052,
"test_loss": 3.8623048021823547,
"test_acc": 23.8,
... | null |
{
"dataset": "cifar100N",
"backbone": "vit_b_16",
"optimizer": "sam",
"epochs": 10,
"milestones": [
7
],
"batch_size": 128,
"lr": 0.01,
"sigma": 0.01,
"beta_coupling": false,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 12,
"data_root": "./data",
"tag": "",
"repo": "txus/fsgld-... | 100 | 73.72 | 72.44 | 71.946 | 784.471455 | 7,842.612702 | [
{
"epoch": 0,
"train_loss": 2.494696511332031,
"train_acc": 49.498,
"test_loss": 1.1563172777996789,
"test_acc": 70.82,
"time": 786.3234913349152
},
{
"epoch": 1,
"train_loss": 1.9306925587032153,
"train_acc": 59.814,
"test_loss": 1.0986020376410666,
"test_acc": 71.96... | null |
{
"dataset": "cifar100N",
"backbone": "vit_b_16",
"optimizer": "sgd",
"epochs": 10,
"milestones": [
7
],
"batch_size": 128,
"lr": 0.01,
"sigma": 0.01,
"beta_coupling": false,
"eta": 0.01,
"rho": 0.05,
"seed": 42,
"workers": 12,
"data_root": "./data",
"tag": "",
"repo": "txus/fsgld-... | 100 | 70.48 | 68.68 | 68.286 | 348.647053 | 3,487.411486 | [
{
"epoch": 0,
"train_loss": 2.317300839802188,
"train_acc": 44.488,
"test_loss": 1.2671460260318805,
"test_acc": 66.55,
"time": 350.08230209350586
},
{
"epoch": 1,
"train_loss": 1.8142727348200804,
"train_acc": 55.796,
"test_loss": 1.1523479442053204,
"test_acc": 69.5... | null |
No dataset card yet
- Downloads last month
- 85